Mixing WordNet, VerbNet and PropBank for studying verb relations

Maria Teresa Pazienza∗, Marco Pennacchiotti ∗, Fabio Massimo Zanzotto†

∗University of Roma Tor Vergata Via del Politecnico 1, Roma, Italy {pazienza, pennacchiotti}@info.uniroma2.it † DISCo, University of Milano Bicocca Via B. Arcimboldi 8 Milano, Italy [email protected] Abstract In this paper we present a novel resource for studying the semantics of verb relations. The resource is created by mixing sense relational knowledge enclosed in WordNet, frame knowledge enclosed in VerbNet and corpus knowledge enclosed in PropBank. As a result, a set of about 1000 frame pairs is made available. A frame pair represents a pair of verbs in a peculiar semantic relation accompanied with specific information, such as: the syntactic-semantic frames of the two verbs, the mapping among their thematic roles and a set of textual examples extracted from the PennTreeBank. We specifically focus on four relation: Troponymy, Causation, Entailment and Antonymy. The different steps required for the mapping are described in detail and statistics on resource mutual coverage are reported. We also propose a practical use of the resource for the task of acquisition and for . A first attempt for automate the mapping among verb arguments is also presented: early experiments shows that simple techniques can achieve good results, up to 85% F-Measure.

1. Introduction scare. Here, the subject of the sentence is the Cause, while The study of verb syntax and semantics is the main focus the object plays the role of the Experiencer and the argu- of many NLP researches, as it is of great help in support- ment introduced by with is an Oblique semantic role. ing a large area of natural language applications (Question Furthermore, to match question and answer, the system Answering, , etc.). should have some knowledge about verb relations (see Fig. Indeed, verbs play a central role in semantics, as they con- 1): specifically, it should know that the verb scare entails vey the core meaning of sentences: the situation described the verb fear (one is scared by something only if he fears in a sentence is in fact expressed through its main verb. it). Finally, the system must know how to match verb argu- All other components of the sentence usually depend and ments. In particular, considering that the Experiencer of the turn around the main verb. Nouns and noun phrases ex- two verbs are the same (”Israel”), and that the Predicate of press the participants to the situation, while adjectives and the question matches the Cause of the answer, it should be other grammatical elements are used to better specify and able to infer that the Theme searched in the question is ac- describe the situation and its participants. tually the Cause in the snippet (”Iran”). The answer is thus To better understand the importance of verbs syntactic and straightforward: semantic behaviour for NLP applications, consider a clas- sical QA system which has to answer the following user ”Israel fears Iran for its nuclear ability.” question: Such kind of linguistic inferences are of great use not only ”What country does Israel fear for its nuclear ability?” in QA, but also in many other NLP tasks (Paraphrasing, , Textual Entailment, etc.) but are A possible answer could be derived through the use of lin- still far from being fully exploited. In fact, even if dedi- guistic knowledge and inference reasoning, from a textual cated syntactic-semantic resources (as VerbNet and Prop- fragment like: Bank for verb structures, WordNet for senses) and ”Iran scares Israel with its nuclear ability.” corpora (e.g., PennTreeBank) are already available , what still lack is a consistent and beneficial integration among In order to carry out such a complex matching, the system them. Specifically, a tight integration between semantic re- should firstly understand the question. It thus must have sources and corpora would allow the use of statistical and some knowledge about the syntactic-semantic behaviour of techniques for supporting tasks on verb the verb fear, that is, the relation between the syntax and semantics. the semantics of the verb arguments. It should then identify In this paper we propose a simple but effective way to inte- the subject of the question as the semantic Experiencer of grate these resources, by creating a mapping among Word- the situation , the object as the Theme, and the argument Net verb relations, VerbNet verb frames and the PennTree- introduced by the preposition in as the semantic Predicate. Bank corpus, using the PropBank database as a bridge be- On the other hand, the system should also recognize the tween the linguistic resources and the corpus. We thus aim syntactic construction of the retrieved snippet governed by to exploit the principle of linking theory, as implemented in

1372 the existing resources, to connect relational verb semantics explicitly linked to the WordNet synset(s) representing its to syntax and finally to surface realizations. sense. As WordNet verb synset lexicon is more fine-grained The goal of our proposed mapping is twofonds. From the than VerbNet classes, more than one synsets can be linked one hand , the link among verb sense relations to their to a verb in a class. In average each verb in VerbNet is frames and eventually to their instances in a corpus can mapped to 3,2 WordNet senses. Moreover, a verb appearing be exploited to carry out complex inference reasoning as in two different classes usually refers to a different sense described in the previous example. From the other hand (and thus links to a different synset). we want to verify the mutual coverage of the different re- VerbNet provides an effective and useful link between the sources, as it is a fair indicator for both estimating the use- syntax and the semantics of a verb. Yet, as the representa- fulness of such kind of mappings and for verifying how lan- tion of complex verb relations and the connection to a ref- guage usage (PropBank) adhere to a specific theory (Verb- erence corpus are not the focus of VerbNet, a mapping to Net). other resources (e.g. WordNet, PropBank) can be usefully The paper is organized as follows. In Sec.2. we briefly de- exploited. scribe the different resources involved in the process. In Sec.3. we outline related works on resource mapping. In 2.3. PropBank and PennTreeBank Sec.4. we then describe our mapping process, together with PropBank (Babko-Malaya et al., 2004) is formed by a verb some statistics on resources coverage. Finally, in Sec.5. we lexicon and a semantically annotated corpus. The lexicon propose an example of effective usage of the mapping for contains about 3600 verbs. Each verb is represented by a Textual Entailment patterns discovery. frame. Each frame is composed by one or more framesets, that refer to specific verb senses. In all, PropBank contains 2. WordNet, VerbNet, PropBank and 5050 framesets. Each frameset is accompanied with its set PennTreeBank of semantic roles (roleset), identified by generic argument 2.1. WordNet labels (Arg0, Arg1, ..., ArgM). The mapping between roles and labels is roleset specific, that is, a label is usually as- WordNet (Miller, 1995) is a lexical database aiming at mod- signed to different roles across the lexicon. elling the human linguistic knowledge in a coherent reposi- Semantic roles in PropBank are more specific than thematic tory inspired by psycholinguistic theories. WordNet covers roles in VerbNet. Indeed, while in VerbNet roles are gen- approximately 150.000 , among nouns, verbs, adjec- eral and valid across different classes, in PropBank they are tives and adverbs. Words are organized in synsets, each strictly tied to a specific roleset. As a consequence, VerbNet representing a lexicalized concept that is linguistically rep- has only 20 thematic roles, while PropBank has more than resented by a set of synonymy words and a gloss describing 1400 roles. Each PropBank roleset is mapped, when possi- the synset itself. WordNet also represents semantic links ble, to a VerbNet class, as it will be described in Sec.4.. among synsets (such as troponymy, entailment, antonymy, As in VerbNet, verb senses in PropBank are fairly coarse- etc.), thus structuring the lexicon in a network of synset re- grained with respect to WordNet, as they are derived by lations. studying the frequency of the verb frame structures in the Verbs are an important part of the network. Roughly 11.000 corpus. Frameset are in fact created grouping those verb verbs are present, divided in 24.632 senses: the polisemy syntactic frames that share the same semantic roles. of verbs is thus quite high, as each verbs has in average 2,3 PropBank frames are used to semantically annotate the senses. Verbs can be linked through specific relations, such Penn Wall Street Journal II with predicate- as troponymy, antonymy, entailment and causation. These argument structures: each occurrence of a verb is properly relations express (directely or indirectely) a lexical entail- annotated together with its argument labels. In all, more ment relation between two verbs (see Sec.5.). than 110.000 PropBank instances are annotated. The fi- Unluckily, as it is mainly devoted to model semantics, nal corpus and the lexicon can be then used as a resource WordNet lacks information about verb syntax, which is for learning useful linguistic phenomena related to verb se- of primary importance in creating a complete resource for mantics. working on verb behaviours. Yet, what is missed in PropBank is a link to WordNet: it is 2.2. VerbNet thus not possible to directely use PropBank knowledge in VerbNet (Kipper et al., 2000) is a hierarchical verb lexicon conjunction with WordNet information on verb relations. based on Levin’s verb classes representing verbs syntactic and semantic information. The main idea is that the syntac- 3. Related Work tic frames of a verb reflect their semantics: verbs sharing The mapping and the integration of different syntactic- similar syntactic behaviour can be then clustered in seman- semantic resources is today a main concern. Many stud- tically coherent classes. Each verb class is constituted by ies have been thus devoted to this problem. In (Kipper et the a set of verbs, their shared syntactic frames, thematic al., 2002), (Kipper et al., 2004) a mapping between Verb- roles and selectional restrictions. Moreover, semantic pred- Net and PropBank is proposed. Finding links between the icates are added to each class to better describe its semantic two resources is not an easy task as it seems. Indeed, Verb- behaviours. Net classes and PropBank framesets have been created us- VerbNet 2.0 contains 237 hierarchically organized classes ing different methodologies: VerbNet is havily based on the (mainly inspired by Levin’s hierarchy) and 5000 verbs. Levin’s classes theory, while PropBank was primarily built Each verb in a class is semantically unambiguous and is looking at verb usage in the PennTree corpus. The mutual

1373 coverage of the two resources is thus a useful means to un- can be learned from the example sentences (see Sec.5.) the derstand to what extent language theory can fit language mapping among the verbs thematic roles (e.g., the Cause of usage in a specific framework. In (Kipper et al., 2004) each scare is mapped to the Theme of fear). PropBank frameset has been manually linked, when possi- The mapping among WordNet, VerbNet and PropBank is ble, to the VerbNet class expressing its syntactic-semantic not straightforward as it seems, for two main reasons: behaviour. Each PropBank role is then mapped to the cor- responding VerbNet thematic roles. • the three resources were built independently and with Not all framesets in PropBank have a link to VerbNet, as a different goals and methodologies. WordNet stems verb sense can be present in PropBank but not in VerbNet. from a psycholinguistic theory and its aim is to model Moreover, not all roles have a mapping to a thematic role. a mental lexicon. VerbNet, as a reflection of Levin’s A PropBank frameset can be mapped to more than one classes, is an instantiation of a linguistic theory: the VerbNet class, as PropBank senses are often more coarse- goal is to model a coherent syntactic-semantic inter- grained. face. PropBank frameset system is a lexicon basically VerbNet coverage seems reasonable: 78.62% of PropBank built from a corpus: it thus stems from language usage sentences in the corpus have an exact matching (all roles and not from language theory. As a gap generally ex- are mapped) to a VerbNet class. Notwithstanding, some ists between language theory and usage, and between VerbNet frames have no corpus instantiations. No mention a psycholinguistic and a purely linguistic representa- is made in the study about mutual coverage of the two lex- tion of lexicon, it is obvious that main discrepancies icon (i.e., the set of frames and classes). can occur in the mapping process. Among other studies, (Shi and Mihalcea, 2005) propose a mapping between VerbNet classes and FrameNet frames • The grain of the resources is different. WordNet is and between VerbNet roles selectional restrictions and spe- fine-grained with respect to verbs and verb senses, cific WordNet semantic classes. The study aims at building while VerbNet and PropBank are both coarse-grained a unified resource to support semantic . but not aligned one with the other. Thus, it is not In (Giuglea and Moschitti, 2004) the link between Verb- always straightforward to obtain a coherent mapping Net and PropBank proposed in (Kipper et al., 2002) is used between verb senses. Moreover, the mismatch be- together with a semi-automatic mapping from VerbNet to tween the grain of VerbNet thematic roles and Prop- FrameNet to improve performance of a role labelling sys- Bank rolesets is another major challenge (as described tem. in (Kipper et al., 2002)). Yet, to our knowledge, no specific research has been de- voted so far to the mapping (and the exploitation) of Word- Net relations, VerbNet classes and PropBank lexicon and All these issues must be taken into account when mapping corpus. the different resources, as described hereafter.

4. Mapping resources 4.1. WordNet to VerbNet Our goal is to link WordNet 2.0 (verb sense relational The mapping between WordNet and VerbNet verb senses knowledge) to VerbNet 2.0 (verb sense frame knowledge) is straightforward, as WordNet senses are explicitly rep- and finally to the PropBank corpus (verb sense frame repos- resented in VerbNet classes. Over the 24.632 verb senses itory). The final objective is thus to have a large set of in WordNet only 4.712 have a correspondence in VerbNet linguistic examples of verb pairs that have some semantic (19.2%). The low semantic coverage of VerbNet is due to relation and specific predicate-argument structures. Once the scope of the two resources: WordNet aim at covering such integrated resource is available, it will be possible to the whole lexicon, while VerbNet looks only at verbs with study and automatically learn how the predicate-argument specific behaviours. On the other hand, all VerbNet senses structures of two verbs are related using the set of corpus are present in WordNet. As VerbNet senses are coarse- examples. grained, they can be mapped to more than one synset. For example in WordNet the verbs scare and fear are in This verbe sense mapping is used to find pairs of VerbNet entailment relation. Specifically, the first sense of scare verb predicate structures corresponding to WordNet verb entails the second sense of fear (scare1 ⇒ fear2). The sense relations. Four type of relations (Miller, 1995) have VerbNet classes corresponding to the two verbs are respec- been studied : troponymy, entailment, causation (gener- tively amuse-31.1 and admire-31.2. It is then possible to ically called lexical entailment relations) ant antonymy. extract the syntax frames for the two verbs, as reported in WordNet accounts for 71.364 verb pairs in these relations Fig.1. Moreover, the two verb classes are mapped in Prop- (e.g., win → compete). Due to the low VerbNet coverage Bank to the frameset scare.01 and fear.01. Thanks to only 11.015 relations (15.4%) can be mapped in VerbNet this mapping it is thus possible to extract from the Prop- (i.e., both verbs are present in VerbNet); the number is fur- Bank corpus all the sentences related to the frameset. At ther reduced to 8.964 when pairs with verbs in the same the end of the mapping process it is thus available a large VerbNet class are discarded. Obviously this is a strong corpus of sentences corresponding to the predicate structure limitation for the final resource, as many relational knowl- allowed for the verb sense relation scare1 ⇒ fear2. Such edge in WordNet is not used. Notwithstanding it is a good a resource can be then refined to find interesting interac- starting point for implementing automatic methodologies tion between the two verb senses structure. For instance, it for mapping.

1374 Figure 1: The frame pair for verbs scare-fear.

4.2. VerbNet to PropBank In all, our resource contains 989 frame pairs (see Table 1). The mapping between VerbNet and PropBank is obtained In average, each frame pair has 50 PennTreeBank example looking at the explicit link in PropBank framesets to Verb- sentences. An example of frame pair is shown in Fig.1 Net classes. Over the 4.498 PropBank framesets (verb VerbNet Frame mapping and Argument mapping are cur- senses), 1.969 have a direct mapping to VerbNet (43,8%), rently on work and done by hand. In Sec.5. possible tech- i.e. they refer to a class in which the verb explicitly ap- niques for automate these processes are presented. pears. On the other hand, 2.363 over 5000 VerbNet verb Many strategies could be applied in the different phases senses are mapped to PropBank (47.3%). Roughly, Prop- to improve the coverage of the final resource. For exam- Bank seems thus to have a better coverage than VerbNet. ple VerbNet verb coverage with respect to WordNet could As a final step thematic roles and PropBank roles are also be improved augmenting classes with new verbs imported manually mapped. from PropBank through an indirect mapping. Many Prop- Bank verbs are in fact linked to VerbNet classes where they 4.3. Frame pair resource are not already present. In such cases, during the building process we discarded the link, as there is not explicit con- Our final resource mapping WordNet, VerbNet and Pr- nection between the new verbs and their WordNet synset. poBank thus consists of a set of frame pairs.A frame pair However we are working at a possible disambiguation pro- is composed by: cess assigning WordNet senses to PropBank verbs, without using VerbNet as a disambiguation resource. • A pair of verb senses in a specific semantic relation; 5. Textual Entailment Patterns Discovery • The syntactic-semantic behaviours of the two verbs, as expresses by VerbNet framesets. A first application of our resource is the discovery of tex- tual entailment patterns. A Textual Entailment patterns, as • A set of mappings between the frames of the two verbs defined in (Dagan and Glickman, 2004), is formed by a text for which the semantic relation holds (VerbNet Frame template T and an hypothesis template H . T and H are mapping); two language expressions such that T entails H, accompa- nied with syntactic properties and possible free slots. For • A mapping between the thematic roles for the two example, the pattern: verbs (Argument mapping); Xsubj scare Yobj Zwith → Ysubj fearXobj Zfor • A set of textual examples for each pair, derived from the PennTreeBank. The actual sentences extracted states that for any syntactically coherent instantiation of X, from the PennTree are assigned to a verb in VerbNet Y and Z, entailment holds. recomposing the sentence using the argument struc- Entailment pattern acquisition is the task of collecting ture of the syntactic frame in the VerbNet class, as these generalized forms, using different techniques rang- done in (Kipper et al., 2004). ing from statistical counts to linguistic analysis. Once a

1375 Relation WordNet WordNet-VerbNet WordNet-VerbNet-PropBank Troponymy 66.756 8.124 888 Entailment 2.251 468 49 Causation 1.278 150 26 Antinomy 1.079 222 26 All 71.364 8.964 989

Table 1: Number of verb relations present in WordNet that can be mapped to VerbNet and successively to PropBank.

large collection of entailment patterns has been acquired, the Patient role is consistent across the two verbs. it can be straightforwardly used to retrieve entailment rela- Yet, more complex methods can be used for supporting the tions in new texts, as needed for specific applications. In simple mapping technique to improve recall, that is partic- such a way, a QA system could map a question like ”What ularly low for causation verbs (only 60%). For example the country does Israel fear for its nuclear ability?” into the H mapping can be boosted by looking at the context of the ar- template Israelsubj fear Yobj nuclear abilityfor. Then, guments: all the examples of each frame are retrieved form all the T templates related to H could be retrieved from the the PennTree corpus, using the mapping between VerbNet collection. Finally, all the answers could be extracted from and PropBank. Once examples for all arguments of the two a corpus, as surface forms matching the T template (e.g. templates are available methods based on Distributional Iransubj scare Israelobj nuclear abilitywith). Hypothesis ((Harris, 1985)) can be applied. As the Distri- Our aim is to use the frame pairs obtained by the mapping butional Hypothesis states that words with similar meaning described in Sec.4. to infer such entailment patterns. In- tend to appear in similar context, we can extend the idea to deed, given two related verb senses, it is possible to retrieve verb arguments. We can state that two matching thematic through VerbNet all their syntactic frames. Once these syn- roles are filled with similar terms. It is then possible to use tactic frames are available, the problem is then to find which co-occurrence vectors (Pantel, 2005) to estimate similarity frames of the two verbs are in entailment relation, and how between two argument contexts. Firstly, co-occurrence vec- the arguments must be matched (VerbNet Frame mapping tor for each argument a of the templates classes are built as and Argument mapping). a list of fillers f accompanied with their frequency or their Argument mapping is in theory a difficult task, as in Verb- pointwise mutual information (miaf ). miaf is defined as Net thematic roles are not meant to be consistent with the follows: syntactic inversion that can take place between two related P (a,f) miaf = log verbs. In particular this assumption is valid for causation P (f)×P (a) verbs. For these class of verbs the subject of the entailed where P (f) is estimated as the frequency of the filler in verb H is usually the object of the entailing verb (e.g. the class, P (a) the frequency of the classes argument and Xsubj raise Yobj → Ysubj rise). In other terms, in most P (a, f) the frequency of the filler in that class argument. cases the subject of T carries out an action that changes The similarity between each argument aT i in T and each ar- the state of the object, that is then described by the verb gument aHj in H is then computed as the distance between H. In such cases, VerbNet roles are not always consistent. the corresponding co-occurrence vectors. The distance For example, in XAgent raise YP atient → YAgent rise, the D(aT i, aHj ) is estimated using cosine similarity (Salton Patient of raise becomes the Agent of rise. and McGill, 1983), as: Yet, we experimented the consistency of VerbNet thematic roles, in order to verify if the argument matching task could P f aT i × aHi be carried out by simply keeping the roles across the rela- D(aT i, aHj) = qP P 2 2 tion (e.g. AgentT → AgentH )(simple mapping). Sur- f aT i × f aT i prisingly, we found that this technique shows good results. We applied this technique by extracting all nouns present We randomly selected 15 verb relations among the lexical in the fillers and using them to build the co-occurrence vec- entailments of our resource and we manually build a gold tors. Early results show a slightly lower precision/recall standard, mapping the arguments of the T verbs to the ar- than the simple mapping technique. guments of the H verbs. We then estimated Precision and Notwithstanding, the two methods can be used together to Recall of the simple mapping technique, by counting how improve recall. We then applied the co-occurrence tech- many roles of the T verb where correctly mapped to the niques to map those arguments that are not matched by roles of the H verb. Simple mapping shows a Precision of simple mapping. On the set of 15 relations, this mixed tech- 96,4% and a Recall of 66,7% (F-measure is 78,7%). Low nique shows an improvement in F-measure from 78,7% to Recall indicates that in many cases an argument of T is not 85,5% (Recall improves from 66,7% to 88%, at cost of a present in H. While results must be taken as only a first lower Precision, 84%). exploratory study, they however seem to outline a certain degree of role consistency among VerbNet classes, even for 6. Conclusions many causation verbs. For example in : In this paper we presented a mapping among WordNet, VerbNet and PropBank for studying semantic verb rela- XExperiencer hurt YP atient → YP atient ache tions. Our main goal was to find the verb frames and textual

1376 examples of verbs in a specific semantic relation stated in Resources and Evaluation (LREC 2004), Lisbon, Portu- WordNet. While the mutual coverage of the resources is not gal. very high, we were able to infer almost 1000 frame pairs. George A. Miller. 1995. WordNet: A lexical database for We believe that this repository could be successfully used English. Communications of the ACM, 38(11):39–41, to support NLP applications, where verb semantics is a con- November. cern, such as in QA, as it has been showed throughout the Patrick Pantel. 2005. Inducing ontological co-occurrence paper. Yet, in order to exploit this knowledge, automatic vectors. In Proceedings of the 43rd Annual Meeting of or semi-automatic techniques are needed to map arguments the Association for Computational Linguistics (ACL’05), across the frames. We showed that such mapping is feasi- pages 125–132, Ann Arbor, Michigan, June. Association ble and that simple techniques boosted with corpus-based for Computational Linguistics. methods should guarantee good performance. However, an G Salton and M J McGill. 1983. Introduction to Modern extensive experiment is still required together with the ap- Information Retrieval. McGraw-Hill. plication of more refined methods based on the Distribu- Lei Shi and Rada Mihalcea. 2005. Putting pieces together: tional Hypothesis. Combining , and for robust se- As a second goal, we verified that the mutual coverage be- mantic parsing. In A. Meyers, editor, 6th International tween VerbNet and PropBank is fairly good. That seems Conference on Computational Linguistics and Intelli- to indicate that Levin’s verb theory instantiated in VerbNet gent (CICLing-05), Mexico City, Mex- has an effective correspondence in language usage, as ex- ico, May 2 - May 7. pressed in the PenntTreeBank corpus. As a future work we are interested in extending the exper- iments for argument mapping and in using different tech- niques for this task, such as those based on the selectional restrictions, as the one proposed in (Shi and Mihalcea, 2005). We also plan to use the discovered frame pairs to infer tex- tual entailment patterns. Once available, these patterns will be used both to give a deeper inside on the textual entail- ment phenomenon and to support QA applications. More- over, the discovered frame pairs represent a valuable re- source that could be used to build Machine Learning mod- els for studying verb relation semantics. 7. References O. Babko-Malaya, M. Palmer, N. Xue, A. Joshi, and S. Kulick. 2004. Proposition bank ii: Delving deeper. In A. Meyers, editor, HLT-NAACL 2004 Workshop: Fron- tiers in Corpus Annotation, pages 17–23, Boston, Mas- sachusetts, USA, May 2 - May 7. Association for Com- putational Linguistics. Ido Dagan and Oren Glickman. 2004. Probabilistic textual entailment: Generic applied modeling of language vari- ability. In Learning Methods for Text Understanding and Mining, Grenoble, France. Ana-Maria Giuglea and Alessandro Moschitti. 2004. Knowledge discovery using framenet, verbnet and prop- bank. In A. Meyers, editor, Workshop on Ontology and Knowledge Discovering at ECML 2004, Pisa, Italy. Zellig Harris. 1985. Distributional structure. In Katz J.J., editor, The Philosophy of Linguistics, pages 26–47. Ox- ford University Press, New York. Karin Kipper, Hoa Trang Dang, and Martha Palmer. 2000. Class-based construction of a verb lexicon. pages 691– 696. Karin Kipper, Martha Palmer, and Owen Rambow. 2002. Extending with verbnet semantic predicates. In AMTA 2002 Workshop: Workshop on Applied Inter- linguas, Tiburon, CA , USA. Karin Kipper, Benjamin Snyder, and Martha Palmer. 2004. Extending a verb-lexicon using a semantically annotated corpus. In urth International Conference on Language

1377